For extensible and accurate machine learned potentials, such as the ANAKIN-ME (ANI) deep learned potential, properly sampling molecular conformational and configurational (chemical) space can be a daunting task. During the development of the ANI-1 data set, an exhaustive sampling of 58k equilibrium molecules with up to 8 total heavy atoms were perturbed along normal modes to generate 17M+ non-equilibrium conformations. With these structures, DFT total energies were calculated to sample the potential surface for training ANI potentials. With the ANI-1 data set and a proper molecular descriptor, ANI potentials have proven to be chemically accurate compared to their reference energies and extensible in their predictive power on molecules that cover similar chemical space as the ANI-1 data set. However, in regions of chemical space sampled poorly in the ANI-1 data set (reaction barriers, non-bonded interactions, and atom dense environments) ANI potentials tend to give less reliable results. For this reason, we develop automated tools for searching and classifying regions of chemical space where an ANI potential produces unreliable results. When poorly sampled regions of chemical space are identified, the algorithm generates new conformations to cover this sparsely sampled region and augments the data set for training ANI potentials. Through this automated and systematic approach of searching chemical space, we develop the ANI-2 data set, which vastly improves the performance of ANI potentials. We show empirical evidence of this performance gain through various test cases. Continuing to improve ANI potentials through smart sampling is shown to be the best way to build a highly efficient, general purpose, and fully-transferable machine learned potential that covers as many important chemical interactions as possible with the goal of maximizing applicability.

Systematic improvement of ANI deep learned potentials through active learning in conformational and configurational space

254th National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.